Lost data neural semantic recovery framework for structural health monitoring based on deep learning

Structural condition perception is a crucial step in structural health monitoring (SHM). Random loss or corruption of sensing data seriously hinders the reliability of the monitoring system. This paper discusses the recovery of randomly lost data in SHM from the perspective of conditional probability generation. A novel data-driven neural semantic recovery framework is proposed, transforming data recovery into a conditional probability modeling problem. This framework uses deep fully convolutional neural networks with an encoder-decoder architecture to capture the overall semantic features of the vibration data, allowing accurate modeling of the behavior of complex conditional probability distributions. Advanced techniques such as dense connections, skip connections, and residual connections significantly improved the network's parameter utilization and recovery performance. Moreover, a novel perceptual loss function is proposed, enabling the network to integrate data loss patterns effectively. The proposed network can be trained end-to-end in a self-supervised manner and perform efficient inferences. Based on the long-term measured acceleration response under the ambient excitation of a pedestrian bridge, the recovery performance and robustness of the model are sufficiently verified and evaluated. The network exhibits excellent recovery accuracy and robustness, even if the loss ratio is as high as 90%. Preliminary evaluation results show that the proposed model can be seamlessly transferred to scenarios with continuous data loss without retraining the network. Finally, the application prospects of the framework in modal identification and anomaly monitoring of structural conditions are demonstrated.